OFM Databank
Building a NSFW AI Model: The ComfyUI + LoRA Pipeline, Explained Honestly

AI & Tools

Building a NSFW AI Model: The ComfyUI + LoRA Pipeline, Explained Honestly

By Yalla Papi · 2026-07-31

Everyone's selling the dream of a fully automated AI model. Here's what the pipeline actually looks like — and where it breaks.

Updated Jul 2026 · sourced from 12 YouTube creators and 9 operator groups

Key takeaways

  • Flux Dev + LoRA (40 images, 4090 or M2) is the current quality benchmark for NSFW generation.
  • Dataset quality kills more LoRAs than bad hyperparameters — 26 clean images beat 100 sloppy ones.
  • ComfyUI is the serious NSFW stack; beginner tools like Higgsfield hit a hard content ceiling fast.
  • Wavespeed and vast.ai undercut RunPod on cost; cloud GPU rentals make the pipeline hardware-agnostic.
  • OnlyFans bans pure-AI models — platform choice is a legal and business decision, not just a tech one.

The Stack Is Real. The Hype Around It Isn't.

Somebody in your feed right now is claiming they generate thousands of images a day from a single selfie, monetize them on OnlyFans, and never touch a camera. Part of that is true.

The part they skip: the dataset prep, the broken LoRA epochs, the platform bans, the $1,600 unban that re-bans in 48 hours. This guide is the part they skip.

We'll walk the full local-generation pipeline — dataset construction, LoRA training, Flux Dev, ComfyUI workflows, face swap, and the cloud shortcuts that actually hold up — with sourced numbers and honest disagreement where it exists.

Step 0: Pick Your Architecture Before You Touch an Image

The base model you train on locks every downstream decision. (Yalla Papi, Jan 2026) Flux Dev has produced the most realistic outputs compared to SDXL, SD 1.5, Chroma, and Z-Image.

Chroma and Z-Image-Turbo show promise but have less track record.

The rule is absolute: the LoRA you train must match the base model you generate with. (Yalla Papi, Jan 2026) A Flux-trained LoRA fed into an SDXL inference environment produces broken output, full stop. (Bjorn Olsen, Dec 2025)

CivitAI itself warns that mismatched base models cause visually "weird and goofy" results.

If you're just starting: train on Flux Dev, generate on Flux Dev. Don't remix the architecture mid-pipeline.

Step 1: Build a Dataset That Won't Waste Three Hours of GPU Time

Bad training images produce a bad model — no hyperparameter tweak fixes this. (Bjorn Olsen, Jan 2026)

The structure that works (vetted):

Target is 26 final images. (Bjorn Olsen, Dec 2025) Train on 30, discard the worst four.

Parameters calibrated for that count will produce off results with a different number. (Bjorn Olsen, Dec 2025)

Consistency details that get skipped:

Include images that show the model's teeth in consistent color and keep pupil color constant across the dataset. (Bjorn Olsen, Dec 2025) Inconsistency in these details causes the LoRA to hallucinate unpredictable eye and tooth appearances in every generation.

Always generate the base anchor image front-facing. (Bjorn Olsen, Jan 2026) A front-on face is easier to lock identity from.

Prompt the base image as a 25-year-old, not 22. (Bjorn Olsen, Dec 2025) Lower epochs underfit and visually regress apparent age — skewing the prompt older compensates.

Removing watermarks before training is not optional. (Bjorn Olsen, Jan 2026) Use Photoshop or any free inpainting tool.

Watermark artifacts bake into the LoRA permanently.

How to Actually Generate That Dataset

Two vetted routes, with different tradeoffs:

Route A — Gemini + Nano Banana Pro: Use Gemini to generate a base set of 26 training images, then use Nano Banana Pro to correct anomalies (mismatching eyes, inconsistent poses) before training. (Bjorn Olsen, Dec 2025) This is described as a significant shortcut over the old Stable Diffusion prompting workflow. (Bjorn Olsen, Jan 2026)

Route B — Civitai cloud training pipeline: Upload images directly to Civitai, select Create → Train a LoRA → Character. (Yalla Papi, Jan 2026) The platform handles training and returns multiple LoRA versions (epochs) to evaluate.

A word of caution from chatter: Multiple operator groups (late 2025 through mid-2026) flagged that Nano Banana Pro embeds Google's SynthID watermark in generated images. At least two separate groups reported AI-model accounts getting flagged or banned because of detectable pixels in NanoBanana-sourced content.

Strip metadata and re-export through CapCut or Photoshop before using these images anywhere public.

Step 2: Training — Parameters, Epochs, and the Traps

Whether you run locally (4090 24GB, M2 Mac with 32GB RAM) or on a cloud pod, the parameters are what move the needle.

Vetted training settings:

LoRA training produces one .safetensors file per epoch. (Bjorn Olsen, Dec 2025) Epochs 3–4 are typically the sweet spot — lower epochs underfit, higher epochs overfit. (Bjorn Olsen, Dec 2025)

Download all five, then compare.

Caution on parameter reuse: Parameters are coupled to a specific image count. (Bjorn Olsen, Dec 2025) The example model trained on 182 images — those exact hyperparameters don't transfer to a 26-image dataset without recalibration.

Trigger word hygiene: Assign a single-word identity token (the model's "name") and prepend it to every image caption during training. (Bjorn Olsen, Dec 2025) If you name the model one thing on CivitAI but trained on a different trigger word, consistency breaks. (Bjorn Olsen, Dec 2025)

Cloud options: RunPod with 100GB of volume storage is the vetted default. (Bjorn Olsen, Dec 2025) Operators in multiple groups (early-to-mid 2026) flag wavespeed.ai and vast.ai as meaningfully cheaper alternatives — wavespeed accepts LoRA URLs directly via API or playground, vast.ai lets you rent bare GPU capacity.

One group described wavespeed as the cheapest current option for NSFW video generation. Worth testing against RunPod costs on your actual workload.

Step 3: Generation — ComfyUI, Samplers, and the Settings That Matter

ComfyUI is the serious stack for NSFW. Full stop.

Multiple operator groups (across at least five distinct groups, Dec 2025–Jun 2026) converged on ComfyUI + custom LoRAs as the highest-quality NSFW image pipeline available locally. One group specifically noted that ComfyUI with LTX 2.3 generates hyper-realistic NSFW content that Higgsfield cannot produce.

Inference settings (vetted for CivitAI generation):

Prompt engineering that actually moves quality:

Reference the model's real physical attributes explicitly in your prompt to compensate for training images that don't capture them accurately. (Yalla Papi, Jan 2026) Stack a secondary "body feature"

LoRA alongside the character LoRA if discrepancies persist. (Yalla Papi, Jan 2026)

Add "realistic skin texture," "candid photo," and "amateur photo" as near-universal prompt modifiers. (Yalla Papi, Feb 2026)

A natural-looking, non-plastic aesthetic consistently outperforms artificial-looking outputs — naturally hanging breasts outperform perky/plastic-looking ones based on documented testing. (Yalla Papi, Feb 2026)

Step 4: Face Swap and the Video Pipeline

The standard content pipeline is images first, video second. (Yalla Papi, Jan 2026) Generate a large batch of images with consistent face and varied prompts, then convert the strongest frames to video.

Starting frames where the subject is already in motion produce better animation than static neutral poses. (Yalla Papi, Feb 2026)

For face swap specifically:

IPAdapter Face ID combined with a LoRA (40 images, 4090 24GB, approximately 3 hours) on Flux Dev is the setup one operator group described for getting a consistent face into existing footage. Build your own ComfyUI face-swap workflow rather than relying on online face-swap services — multiple groups flagged online services as unreliable for video.

One group specifically noted that FaceFusion and Akool both work for image swaps, but video face swap in ComfyUI is a different, harder problem.

Patryk's approach is worth noting: use Higgsfield's face-swap tool to overlay the model's face onto videos you record yourself, or onto clips from other creators. (Patryk, Feb 2026) Two methods — record yourself doing an action and swap the face in, or take a viral reel and replace the face.

This is the hybrid approach and currently produces more convincing results than pure generation for many operators. (Patryk, Mar 2026)

Where Operators Disagree — Both Sides, Plainly

ComfyUI vs. simpler tools: Multiple groups position ComfyUI as the only real NSFW option. But at least one group argues ComfyUI is poor unless self-hosted, and that for non-NSFW content, Kling is both cheaper and better.

Beginners are advised in one group to start with Kling via Higgsfield and only move to ComfyUI when NSFW is the explicit goal.

Higgsfield censorship: Groups conflict sharply here. Some describe Higgsfield's NSFW filter as a hard wall that makes it useless for adult content.

Others report that Seedream 4.5 inside Higgsfield on the creator plan (~$150/month) generates "endless NSFW" content. One operator claimed $5k in revenue since using this setup.

This cannot be reconciled from available evidence — results may depend on account tier, region, or content specifics.

Nano Banana quality vs. censorship: Several groups rate Nano Banana Pro as the highest-quality consistency tool while simultaneously warning it's strict on NSFW/bikini prompts and embeds detectable SynthID watermarks. Seedream 4.5 is positioned as less restricted but lower quality.

At least two groups say Seedream 4.0 bypasses NSFW restrictions that 4.5 enforces. No consensus.

Platform viability — this is the one that matters most: At least four separate groups across multiple months agree: OnlyFans bans pure-AI model accounts, and OF will ban you at the first mistake. Fanvue explicitly allows AI models, requires you to select "AI" during KYC, and verify with your own manager ID.

However, one group warns that Fanvue flags AI models that appear too young and resets earnings if content is flagged at review. One group notes that being on Fanvue signals to audiences that the creator is AI-generated — a conversion consideration.

MYM is also cited as permitting AI models.

The Cheapest Viable Cloud Stack Right Now

You do not need an H100 ($30,000, roughly five RTX 5090s) to run this pipeline. (Yalla Papi, Jan 2026) You need access to one.

  • RunPod: Vetted default for LoRA training. Allocate 100GB volume storage. (Bjorn Olsen, Dec 2025)
  • Wavespeed.ai: Flagged by multiple operator groups (early-to-mid 2026) as the cheapest option for NSFW video generation; accepts LoRA URLs via API.
  • vast.ai: Bare GPU rental; useful when RunPod queues are long.
  • SwarmUI on cloud GPU: Full AI content pipeline accessible without local hardware. (Yalla Papi, Jan 2026) Setup takes roughly 20 minutes once learned.

The local alternative — M2 Mac with 32GB RAM or a 4090 — is viable for LoRA training and image generation. NSFW video at quality is a harder problem locally; one group noted that even with 96GB VRAM, NSFW video still fails AI-detection checks.

Bottom Line

The pipeline is real and it works. The gap between "works" and "scales profitably" is almost entirely dataset quality and platform discipline.

Clean 26-image dataset. Flux Dev base model.

Five training epochs — pick epoch 3 or 4. ComfyUI for NSFW generation with CFG 5.5 and DPM++ 2M Karras.

Images first, video second. Wavespeed or vast.ai if you don't want to own hardware.

Then make the platform decision with your eyes open: OnlyFans for hybrid (real-girl verified, AI-enhanced content), Fanvue for pure-AI with disclosure. Getting this backwards is how you lose a built account.

Sources

On the record (YouTube creators):

  • Yalla Papi — How I Make AI Content From Real Models (LORA TUTORIAL), Jan 2026. Watch ↗
  • Bjorn Olsen — High-Quality AI OFM Models Using Gemini 3 + LoRA Training Method, Dec 2025. Watch ↗
  • Yalla Papi — How I Make 100+ AI videos of my OnlyFans creators every day, Jan 2026. Watch ↗
  • Bjorn Olsen — High Consistency LoRA Training Method for AI OFM Model (Step-by-Step), Dec 2025. Watch ↗
  • Bjorn Olsen — How to Create High Consistency NSFW Content for your AI OFM model (CivitAI Guide), Dec 2025. Watch ↗
  • Bjorn Olsen — SECRET Lora Training Hack for ULTRA REALISTIC AI OFM Model, Dec 2025. Watch ↗
  • Bjorn Olsen — Fix AI Model Face Consistency Forever (AI OFM Method), Jan 2026. Watch ↗
  • Yalla Papi — My Unreasonable OFM Predictions for 2026, Jan 2026. Watch ↗
  • Yalla Papi — Launching an AI OF model from scratch in 59 minutes, Feb 2026. Watch ↗
  • Patryk — How AI can make you $50k/month (OFM), Feb 2026. Watch ↗
  • Patryk — AI could make you rich but you're using it wrong..., Mar 2026. Watch ↗

Community intelligence: 172 operator claims aggregated from 9 separate private OFM groups (Dec 2025–Jun 2026), corroboration counted across groups. Group identities are withheld to protect sources; browse the underlying intel in the Community Intel Wiki.